1012 IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS, VOL. 26, NO.

1, JANUARY 2020

The Perceptual Proxies of Visual Comparison

Nicole Jardine, Brian D. Ondov, Niklas Elmqvist, Senior Member, IEEE, Steven Franconeri




 
 

 
 
 
 

! 
 

 ! 
 



 
  
 
 

 "
 

 # 




 




 
 
 

 # 
 


 

 
 

 
 
  
 
 
 #   

     

  
   
"
 
 

  
 
 

Fig. 1: What’s visual in visual comparisons, such as finding the larger mean value? We identify mark arrangements that allow
for better performance across comparison tasks. Combining previous results with the results of two new tasks fails to produce a
clean ranking of arrangement effectiveness across tasks. We argue that to explain these complex patterns of performance, we first
need a perceptual explanation of how visual comparison actually unfolds. Viewers likely perform these mathematical comparison
operations with perceptual proxies. We propose and evaluate a candidate set of proxies for two visual comparison tasks.

Abstract—Perceptual tasks in visualizations often involve comparisons. Of two sets of values depicted in two charts, which set had
values that were the highest overall? Which had the widest range? Prior empirical work found that the performance on different
visual comparison tasks (e.g., “biggest delta”, “biggest correlation”) varied widely across different combinations of marks and spatial
arrangements. In this paper, we expand upon these combinations in an empirical evaluation of two new comparison tasks: the
“biggest mean” and “biggest range” between two sets of values. We used a staircase procedure to titrate the difficulty of the data
comparison to assess which arrangements produced the most precise comparisons for each task. We find visual comparisons of
biggest mean and biggest range are supported by some chart arrangements more than others, and that this pattern is substantially
different from the pattern for other tasks. To synthesize these dissonant findings, we argue that we must understand which features
of a visualization are actually used by the human visual system to solve a given task. We call these perceptual proxies. For example,
when comparing the means of two bar charts, the visual system might use a “Mean length” proxy that isolates the actual lengths of the
bars and then constructs a true average across these lengths. Alternatively, it might use a “Hull Area” proxy that perceives an implied
hull bounded by the bars of each chart and then compares the areas of these hulls. We propose a series of potential proxies across
different tasks, marks, and spatial arrangements. Simple models of these proxies can be empirically evaluated for their explanatory
power by matching their performance to human performance across these marks, arrangements, and tasks. We use this process to
highlight candidates for perceptual proxies that might scale more broadly to explain performance in visual comparison.
Index Terms—Graphical perception, visual perception, visual comparison, crowdsourced evaluation

1 I NTRODUCTION
Visual comparison is a core perceptual task in data visualizations [10].
An epidemiologist might use two bar charts to assess whether, across
all age groups, there is a larger overall population of women than men.
An education researcher might use two bar charts to assess whether
one group of students’ test scores has a larger range than another. Nei-

Nicole Jardine is affilitated with Northwestern University and with the
Cook County Assessor’s Office in Chicago, IL, USA. E-mail:
njardine@cookcountyassessor.com.

Brian Ondov is with the National Institutes of Health in Bethesda, MD,
USA and University of Maryland in College Park, MD, USA. E-mail:
ondovb@umd.edu.

Steven Franconeri is with Northwestern University in Evanston, IL, USA.
E-mail: franconeri@northwestern.edu.

Niklas Elmqvist is with University of Maryland in College Park, MD, USA.
E-mail: elm@umd.edu.
Manuscript received 31 Mar. 2019; accepted 1 Aug. 2019.
Date of publication 16 Aug. 2019; date of current version 20 Oct. 2019.
For information on obtaining reprints of this article, please send e-mail to:
reprints@ieee.org, and reference the Digital Object Identifier below.
Digital Object Identifier no. 10.1109/TVCG.2019.2934786
ther of these comparison tasks need rely on the averages of each of
these sets, or identification of individual values. They simply require
a judgment of which set’s mean or spread is larger than the other. Al-
though certain visual channels [3, 18] are known to convey higher-
precision information (position) than other marks (hue), existing eval-
uations of visual comparison performance suggest that there is no sin-
gle mark or spatial arrangement that optimizes visual comparison.
In prior work, Ondov et al. [20] evaluated the perceptual preci-
sion of visual comparisons of the “biggest delta between items” and
“biggest correlation between sets” for different visualization marks
(bar, line) and spatial arrangements (stacked, mirrored, adjacent, su-
perposed, and animated). Precision was not optimized by a single
mark type or spatial arrangement. Instead, the precision of visual
comparison depended on an interaction of mark, arrangement, and
task (Figure 2). The best static chart for a precise delta compari-
son, for example, was one that was spatially superposed, rather than
juxtaposed in a stacked format (Figure 1), validating an intuitively-

1077-2626 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
JARDINE ET AL.: THE PERCEPTUAL PROXIES OF VISUAL COMPARISON
Fig. 2: Visual comparison depends not on a single dimension of mark,
arrangement, or task, but of the interactions between them. These
interactions can be represented as a cube. Our present goal is not to
examine the full space of the cube, but rather to understand how a
viewer uses visual features to serve analytic task goals depending on
the marks and arrangements they see.
motivated guideline from Gleicher et al. [10]. Not predicted by current
guidelines was the discovery that, to support biggest delta comparisons
in data, animation provided the most precise performance. Animation,
however, did not perform as well for comparisons of correlations.
We first extend this work to empirically evaluate performance
across these arrangements for two new tasks—“biggest mean” and
“biggest range”—and again find that performance is strongly impacted
by spatial arrangement. Comparison was most precise when these two
datasets were vertically stacked, and least precise when the datasets
were superposed. This pattern of which arrangements were best was
strikingly different than for the previous pattern for “biggest delta be-
tween items” and “biggest correlation between sets.”
Why is there not a single clean emerging answer, where a given ar-
rangement is best across various tasks? This empirical evidence for the
more complex nature of visual comparison is consistent with the idea
that it requires a series of visual actions at a variety of scales from one
object, to multiple objects, to whole sets of objects [9]. Taxonomies
of visual comparison describe multiple stages of perceptual and cog-
nitive steps [4, 9], and vary in describing one or many types of visual
comparisons, but the visual mechanisms supporting these stages are
unclear. We argue that an empirical description of the precision of vi-
sual comparison across each combination of mark × arrangement ×
task would be valuable, but unlikely to scale to have predictive be-
yond its status as a lookup table. A different approach is required. We
propose that instead of continuing to fill out the entries of the cube in
Figure 2, it may be more productive to study perceptual proxies of a vi-
sualization are actually used to reason about a visual comparison task.
The goal of this approach is to begin to identify the proxies for visual
comparison, as opposed to merely gathering additional empirical data.
We propose several candidate visual proxies and implement them
as simulations. This allows us to evaluate each proxy’s objective per-
formance in performing the same task given to the human participants.
But by comparing the performance of these proxies to human perfor-
mance on the same questions, we can rank proxies by which most
closely mirrors human performance.
Our evaluation of perceptual proxies suggests that although these
two comparison tasks show similar arrangement-driven patterns of re-
sults, these patterns are consistent with different proxies. To compare
means, the visual features that best predict human performance are the
ensembles of lengths and the centroids of the bars. To compare ranges,
the best-predictive visual features were those that compared deltas be-
1013
tween all items within a set or only between neighboring items.
The complex dependency of visual comparison performance on
combinations of marks, arrangements, and tasks might soon be pre-
dicted by a model that accounts for such a set of perceptual proxies.
We speculate that these proxies follow two broad categories of global
or set-based visual comparisons and focal or item-based visual com-
parisons, drawing from perceptual psychology research on that divi-
sion in types of visual processing.
This work contributes (1) results on how visualization design ar-
rangement affects two new comparison tasks with two new crowd-
sourced graphical perception experiments for visual comparisons of
“biggest mean” and “biggest range”, (2) a framework of perceptual
proxies for visual comparison, (3) implementations of these proxies
for empirical evaluation, and (4) data generation procedures designed
to estimate the amount of signal needed in the data to support a given
visual comparison between sets of items. Our findings present a first
step toward a model of human perception during visual comparison.
2 R ELATED W ORK
Here we review empirical research on visual comparison and suggest
that visual comparisons can often be classified as being made between
isolated parts (i.e., a bar distinct from other bars in its set) or whole sets
(i.e., all the bars). Frequently, these correspond to analytic tasks for
which a goal is identification or comparison of items, or of sets, in data.
We propose that these analytic task goals correspond to proxies that
determine the visual features that a viewer uses for visual comparison.
2.1 Visual Comparison
Frameworks of visual comparison are often driven by the analytic
goals of the viewer; for example, Amar et al. [2] names comparison as
a high-level “compound task” central to many specific analysis tasks.
Gleicher et al. [9] conducted a review of the taxonomies of visual com-
parison with the goal of a top-down approach examining what people
do when they do visual comparison. They propose that, broadly, to
“compare” in a visualization involves at minimum three components:
targets (which multiple “items” are being compared), the relationship
between these items, and an action (a mechanism operating on the re-
lationship between these targets). These targets may correspond to
items or sets of data. Yi et al. [28] discussed adding a comparison
task to their seven-task taxonomy, but ultimately decided against it be-
cause “compare is a higher-level user goal or objective.” In contrast,
our work here and in past work [20] frames comparison as a relatively
low-level perceptual task.
2.2 Visual Comparison: Parts vs. Wholes
Visual comparison across multiple series can be comparisons of items-
to-items (focal) or sets-to-sets (global). We refer to these as different
visual spans of comparison. One study tested visual comparisons be-
tween smaller regions of time series charts, or between larger areas
of time series charts [14]. The data were consistent with the idea that
these are distinct visual actions: viewers conducting focal visual com-
parisons of small regions benefited from shared-space charts that over-
lapped in space, whereas viewers conducting visual comparisons over
the entire sets were better served by separated-space charts.
Another study [20] investigated the perceptual precision of two vi-
sual tasks, and tested a series of chart arrangements to see which ar-
rangements best supported each visual comparison. In one task, peo-
ple saw two sets of 7 items and compared them to identify which of the
7 items changed the most between the first and second chart. In this
item-to-item comparison, visual comparison was most precise when
the charts were spatially superposed within a single chart space, or had
an animated transition between them (temporally superposed). These
arrangements created visually salient features that mapped to these
item-to-item changes, either as overhang in superposed charts or as a
salient motion cue that captured attention in the animated charts. Com-
parison was less precise when the sets were separated spatially such
that the two chart axes were arranged horizontally, vertically, or mir-
roring each other. In another task, people saw two pairs of bar charts
that were correlated with each other to some degree. The viewer’s task
1014 IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS, VOL. 26,
was to pick the pair with the strongest correlation. Here, performance
was best when each pair of correlated charts had axes that mirrored
each other. Speculatively, viewers could rely on he rapid perception of
symmetry between chart items, which in this task happened to indicate
correlation between charts.
These focal vs. global modes have analogues in perceptual psychol-
ogy research as different perceptual modes. A viewer can opt between
these modes (with a mandatory dominance for global mode for a first
glance at a new image) to flexibly meet the demands of focal tasks
(identify items within sets of larger items) or global tasks. In other
words, when people are presented with a visual stimulus, they can
flexibly choose whether to attend to the “forest” or the “trees” of that
stimulus at varying spatial scales. These different attentional modes
switch which scope of a visual stimulus is used for a task goal [19].
In sum, we predict that item-to-item comparisons are facilitated by
animated and superposed charts that place these items in close prox-
imity to each other, allowing a focused mode to subserve this more fo-
cused task. Set-to-set comparisons may be facilitated by arrangements
that spatially separate these sets, allowing the visual system treat each
set globally as its own unit, because preserving the values of specific
items is not necessary. Different arrangements support different com-
parisons because, we propose, these arrangements offer visual proxies
that people actually use to conduct visual comparisons. Biggest delta
benefits from animation because a salient visual cue (motion speed of
an item changing the most) naturally maps to, or is compatible with,
the apprehension that the data value is also changing the most.
2.3 Interim Summary
Empirical evaluation of how human observers perceive values or rela-
tionships in a visualization suggests that people can rely on a variety of
proxies, operating over the marks and arrangements and other visual
properties of a visualization, to meet the demands of an analytic task.
3 H YPOTHESES
Our original hypothesis was that the two new tasks MaxMean and
MaxRange would show similar patterns of performance to the Max-
Correlation and MaxDelta tasks from previous work [20]: that com-
parisons of means would, like comparison of correlation, benefit from
mirrored charts. We were surprised to see an almost opposite pattern of
performance, leading us to turn out attention toward the path of seek-
ing perceptual proxies that might provide more explanatory power for
why these strikingly different patterns emerged.
4 M ETHODS
Our first goal is to quantify the perceptual precision with which human
observers can perform visual comparison between two charts of hori-
zontal bars, and to measure whether that precision differs based on the
spatial arrangement of those charts.
4.1 Tasks
We chose two tasks to build on previous work:

M AX M EAN: Of two sets, which had the largest average (mean)
value? Difficulty is increased by reducing the delta between
the mean values, so that the difference between sets is less
distinguishable. Displays were controlled so that the largest
single-item in a chart was not predictive of that chart having the
largest mean and so that charts in a trial were of approximately
equal variance. Within-chart variance ranged from .04 to .09.
Harder discriminations (smaller mean deltas) spanned the low
to high variance range, whereas easier discriminations tended to
be lower variance. The two extreme values that bounded data
generation were directly included in a randomly selected chart,
ensuring that the highest or lowest individual value did not corre-
late with the correct answer, and allowing us to examine whether
participants used this as a proxy.

M AX R ANGE: Of two sets, which had the widest range between
its min and max values? Difficulty is adjusted by varying the
NO. 1, JANUARY 2020
delta value between the range widths of the two charts. Since
range may be a less widely-understood concept, we gave our
participants a detailed description with a simple example, both at
the start of the trials and each time they were incorrect in training
trials. See past studies [12, 22] for similar tasks.
4.2 Titer Staircase Method
Our goal was to quantify the magnitude of the difference of means
for the M AX M EAN task, and the magnitude of the difference of range
widths for the M AX R ANGE task, for each arrangement. We dynami-
cally titrated stimulus difficulty using a staircase method [20]. Briefly,
this method scales task difficulty on a trial-by-trial difference. The end
result quantifies a titer: a value between 0 and 1 that scales the mag-
nitude of the difference between stimuli to determine the threshold at
which a participant can barely perform a discrimination task (expected
performance of 75%). Fig. 3 illustrates this.
In the M AX M EAN task, the titer controlled whether there was a
large difference between the two chart means (large titer: easy task) or
a smaller difference between the two means (smaller titer: harder task).
For M AX R ANGE, the titer controlled whether there was a large differ-
ence between the range widths of the two chart means or a smaller
difference between the range widths of the two chart means (Fig. 1,
left; orange), making this pair stand out more from the baseline pair.
Titers and stimulus datasets changed trial-by-trial depending on par-
ticipant performance for the previous trial. Briefly, the initial titer
value that scaled the difference-of-means or the difference-of-ranges
was 0.5. Depending on whether the participant’s discrimination in that
trial was correct or incorrect, the next trial adjusted the titer by -0.01
or +0.03. The goal of this staircase procedure is that by the end of
the trials, the titer reflects a stable magnitude of signal that allows the
participant to perform with 75% accuracy for that arrangement.
4.3 Arrangements
As in previous work [20], datasets were presented in blocks of 5 dif-
ferent arrangements (Fig. 1):

Stacked: Vertically juxtaposed small multiples (i.e. one chart is
placed above the other).

Adjacent: A more commonly used instance of small multiples,
in which data series are horizontally juxtaposed.

Mirrored: This horizontally “mirrored” variation of adjacent
opposes the direction of the x-axis in each chart. The Gestalt
nature of bilateral symmetry suggests this arrangement prompts
“set” proxies rather than “item” proxies in viewers.

Superposed: A combined chart depicting both data series within
the same space. Past work has claimed that overlaying values, or
superposition, minimizes eye movements and memory load, and
may lead to efficient comparison [10].

Animated: In this “arrangement,” a single chart is transitioned,
or morphed, from one data series to another over time, using
cubic interpolation to ease the transitions [7].
In trials, the order of these 5 blocks was rotated, and each rotation
reversed, for a total of 10 possible orderings, each of which was pre-
sented to 5 participants.
4.4 Task and Procedure
Before each trial began, the screen contained a centrally placed fixa-
tion dot and outlines of where the charts would appear. Participants
clicked a button to start the trial. After a countdown, the visualiza-
tion appeared for a short, fixed time. Static and animated charts were
shown for 1.5 seconds. In contrast to previous work, at the end of the
impression, both sets of data were removed from the display. Partici-
pants then clicked on the orange or blue button corresponding to the or-
ange or blue set of bars to provide a response. For the M AX M EAN task
they were instructed to “Click on the chart that had the biggest mean
JARDINE ET AL.: THE PERCEPTUAL PROXIES OF VISUAL COMPARISON

 
 

  
 

 
 
Fig. 3: In the staircase procedure, a correct response produces a smaller difference in the subsequent trial.
values”; for M AX R ANGE to “Click on the chart that had the widest
range between min and max values.” Participants were informed if
they were correct and, if incorrect, what the correct answer was. This
feedback was provided to make the task more engaging and to rein-
force the goal. Between trials, the titer was adjusted based on the
response. To seek 75% accuracy during trials, the titer was increased
three times as much for an incorrect answer as it was decreased for
a correct answer (see Figure 3). Dynamic data generation according
to the titer value is described in Supplemental Materials A. Each ex-
periment included all five arrangements. There were twenty trials for
each arrangement, and arrangements were blocked. The order of the
arrangement blocks was changed between participants.
4.5 Training
Before training, participants were shown examples of stimuli and the
task. Before each arrangement block, participants were given a time-
unconstrained version of the task, which they were required to answer
correctly before proceeding (once for the M AX M EAN task, three times
for M AX R ANGE). Additionally, the first non-animated arrangement
given to a participant followed untimed training with three timed train-
ing trials, which were identical to the real trials except that they always
had the easiest (largest) titer. Data were regenerated on incorrectly an-
swered training answers to minimize answering by elimination. Video
demos are in Supplemental Materials.
4.6 Participant Recruitment
Based on previous work, we predicted N = 50 would provide suffi-
cient statistical power to reliably detect the presence or absence of
an effect of arrangement. We also expected that more participants
would struggle to understand the MaxRange task, so we collected data
from 50 MTurk workers for the MaxMean task and 54 MTurk work-
ers for MaxRange. Participants were asked to self-select out of the
study if they had color vision deficiencies. Each participant com-
pleted the staircase procedure for all 5 arrangements of one of the
tasks (M AX M EAN or M AX R ANGE). Worker IDs were used to ensure
uniqueness of participants across all such combinations. All workers
recruited for participation were adults in the United States.
5 R ESULTS
We evaluated the magnitude of deltas required in the data for non-
expert participants to reliably identify the set with the largest mean
(Experiment 1) or the largest range (Experiment 2).
5.1 Exclusion Criterion
The dependent measure in these experiments is the average titer value
from the final 10 trials of each task. We excluded datasets from par-
ticipants whose average titers (averaged over spatial arrangement) was
more than 2 standard deviations from the mean. This procedure elim-
inated 1 observer each from the M AX M EAN and M AX R ANGE tasks.
We also adopted a second criterion. In a staircase procedure, the
goal is to find a converged titer value for which a participant is 75%
accurate. The procedure fails if a participant repeatedly reaches ceiling
performance (a minimum titer value of 0.01) or floor performance (the
1015

 

 
  
 
  
 

  
 
maximum titer value of 1.0) because at this point the stimuli cannot
titrate difficulty beyond these floors and ceilings.
Because viewers performed tasks for 5 arrangements, we excluded
participants for whom there were at least 5 trials of floor or ceiling
titer values. These criteria excluded 0 from the M AX M EAN task, but
for M AX R ANGE there was 1 trial in which a participant reached ceil-
ing performance and 109 trials who repeatedly reached the floor titer
(largest delta). We excluded 7 participants for whom there were at
least 5 (up to 22) trials of floor titer values (one of whom was also the
participant excluded with the standard deviation procedure), leaving
N = 49 for the M AX M EAN task and N = 47 for M AX R ANGE.
5.2 Titer Analysis
We computed each observer’s mean titer values from the final 10 trials
for each arrangement. We used the final 10 trials because visual evalu-
ation of trial-by-trial data suggested that this was approximately when
the staircase procedure stabilized around a narrow range of titers, for
most participants. Thus we analyze the final 10 titer values achieved
for each of the five arrangements, for each subject.
5.3 Exp. 1 and 2: MaxMean and MaxRange
Figure 4 (far left) displays the mean final 10 delta values for the
M AX M EAN task, and (second from the left) displays the mean values
for the M AX R ANGE task. These titer values correspond to the differ-
ences between the charts being compared. Means could be discrimi-
nated when they differed by approximately 5-8% of the chart axis, and
range widths when they differed by approximately 14-17%. For both
tasks, the precision of visual comparison was affected by arrangement.
Titer values for the present experiment were analyzed with a mixed
ANOVA to test for experiment-level and arrangement-level effects.
Titer values varied between experiments, F(1, 94) = 9.06, p = .003,
η p2 = 0.09, but this is likely because the titer values scale to different
stimulus changes between the two experiments. As such we avoid a
meaningful comparison between differing titer values.
More meaningful is that there was a significant effect of arrange-
ment on precision, F(3.09, 290.7) = 8.17, p < .0001, η p2 = 0.08, with-
out evidence for an interaction between arrangement and experiment,
F(3.09, 290.7) = .34, p = .85, η p2 = 0.004, both Greenhouse-Geisser
corrected. This suggests that arrangement produces largely similar ef-
fects on the precision of visual comparisons of means and of ranges.
We conducted pairwise comparisons within each experiment.
Stacked charts were the most precise arrangement overall for both
tasks. Precision was better in stacked relative to adjacent charts for
the M AX R ANGE task, t(49) = 2.73, p = .009, although not signifi-
cantly better in the M AX R ANGE task, t(47) = 1.70, p = .09. Super-
posed charts resulted in the lowest precision for both tasks. Note that
these patterns are strikingly different compared to prior evaluation of
visual comparisons of items, which were best supported by animated
and superposed charts.
5.4 Accuracy
The goal of a staircase procedure is to titrate the task’s difficulty so dif-
ficulty might change across arrangements, but that accuracy is equiv-
1016 IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS, VOL. 26,
  

  



 
$%

&%
      !
 
Fig. 4: Means of averaged final titer values across participants performing the M AX M EAN and M AX R ANGE tasks. Smaller titers correspond to
more precise differences between means (range widths). The precision of both the M AX M EAN and M AX R ANGE tasks was affected by chart
arrangements. Also presented are titer values from previously published empirical evaluations of the precision of other comparison tasks. Note
that different chart arrangements support different visual comparisons. Gray bars represent 95% confidence intervals.
alent between arrangements. Mean accuracy in the M AX M EAN task
for each arrangement ranged from 76.4% (stacked) to 79.9% (mirror),
with no evidence that accuracy was different between arrangements.
This suggests the staircase procedure reliably converged for this task.
Mean accuracy in the M AX R ANGE task ranged from 75.2% (super-
posed) to 84.7% (stacked), and a repeated measures ANOVA found
that accuracy consistently differed between arrangements, F(4, 184) =
4.34, p = .002. The staircase procedure did not reliably converge
for all arrangements in the M AX R ANGE task due to large effects of
arrangements on people’s ability to perceive range widths. Stacked
charts allowed for higher accuracy and high precision than other ar-
rangements. In a pilot version of this experiment with fewer partici-
pants, we tested a larger initial titer value so that participants unfamil-
iar with statistical ranges could use a very large signal in this task, but
found the same pattern: superposed charts simply outperform stacked
charts regardless of initial task difficulty. Because of these strong ef-
fects of chart arrangement, future work using this method might better
titrate task difficulty by relying on a greater number trials, or a differ-
ent kind of adaptive titration procedure.
6 D ISCUSSION
The precision of visual comparison of M AX M EAN and M AX R ANGE
tasks were best supported by vertically stacked charts, and least sup-
ported by superimposed charts. This is in contrast to previous findings
for item-item comparisons, which were best supported by animated
and superimposed charts. These differential findings are consistent
with the principle that item-item and set-set comparisons of data are
supported by arrangements that enable focal and global visual feature
comparison, respectively. That vertically stacked charts were best may
be unsurprising because marks are horizontally extending bars. View-
ers can simply slice downward to extract lengths between charts.
We hypothesize that the reason these arrangements best supported
these tasks is that visual system can, somewhat flexibly, adopt a series
of focal or global “perceptual proxies” that operate on a visualization.
To understand visual comparison we must understand not only what
visual features exist in a given visualization, but when they are used
for which tasks. But these are post-hoc explanations. Are people actu-
ally using that proxy for both tasks? Or are people using different vi-
sual features to support M AX M EAN and M AX R ANGE? This question
cannot be answered by evaluating the precision of visual comparisons.
Next we study whether the same patterns of precision arising from the
same arrangements for different tasks arise because of the different
visual features that people use for comparison.
NO. 1, JANUARY 2020
    
   
"   " #   " #
 
7 P ERCEPTUAL P ROXIES
Instead of mathematical operations, people more likely rely on heuris-
tic perceptual proxies to extract data values and patterns from data vi-
sualizations. Heuristics are shortcuts that rely on a simplified metric—
a proxy metric—to convey the desired information. Perceptual heuris-
tics are easily computable features that (at least) correlate with the
right answer, allowing viewers of visualizations to use visual features
as accurate or inaccurate proxies for the data those features represent.
One example is the perception of correlation in scatterplots. The
perceptual process does not appear to calculate the true mathematical
correlation, and there are instead proposals for multiple proxies that
might underlie correlation perception [12, 22, 27], including the as-
pect ratio of the bounding box surrounding the points [27]. This proxy
can be efficient because it relies on a rapid perceptual process of in-
specting a shape boundary around the points. This proxy is also fairly
accurate [27]: the width of the bounding box of the visualization cor-
responds strongly to the correlation in the data.
Scatterplots are commonly used to show correlation data, but not all
links between visualization and analytic task are so strong. Further-
more, an analyst might not always know what kind of visualization
they will see. Finally, the selection of a proxy will be strongly affected
by the visual features that are available in a visualization.
Different proxies may afford not only different data patterns, but
different conceptual associations of what those values might mean.
The same two data points graphed as two bars or as two endpoints of a
line chart can evoke different visual actions taken on visual features of
the visualization. Zacks and Tversky [29] presented simple line or bar
charts to participants for open description. Participants’ descriptions
of bar charts overwhelmingly tended to involve discretizing words,
such as “Y is higher than Z,” and descriptions of line charts entirely
used continuous relations, such as “as X increases, Y decreases.” This
bar-line message correspondence seems to occur because the type of
mark is associated with metaphors of bars being containers or groups,
in contrast to lines, which are continuous entities.
7.1 Candidate Proxies
A visualization contains any number of visual features potentially
available as a proxy for a given task, such as the lengths of the top
most items of each set, or the perceived symmetry of each set. Differ-
ent visual features might be better proxies than others for different vi-
sual comparisons. Here we explore which visual features appear to be
most similar to participant performance (making the same decision),
when used as a proxy for M AX M EAN or M AX R ANGE. We developed
two broad categories of candidate features, informed by research in
both visualization and perceptual psychology.
JARDINE ET AL.: THE PERCEPTUAL PROXIES OF VISUAL COMPARISON 1017

033*#-&
1208*&3 &3$2*14*0/ *35"-
$0(/*4*0/
12*/$*1-&

&"/ 842"$43
-&/(4)3
0'
#"23
0'
&"$)
3&4
&01-&
$"/
&842"$4
4)&
.&"/
3*:&
0'
"
3&4

$0.154&3
&/3&.#-&3
$)003&3
$)"24
0'
*4&.3
!
4)05()
4)"4
.&"/
*3
-*,&-9

7*4)
-0/(&2
&/3&.#-& 4)205()
30.&
12089

&/420*% *$,3
$)"24
7*4)
-"2(&34
$&/420*%
0'
4)&
9&
.06&.&/43
2"1*%-9
%&1-09
40

#"2
"2&"3
"-0/(
+534
2&-&6"/4
8
"8*3 $&/420*%3
0'
(20513
#54
4)03&
$&/420*%3

"11&"2
40
#&
$0.154&%
"$2033
4)&

#05/%*/(
)5--
0'
4)&
0#+&$43
/04
4)&*2
425&

5--
2&" "-$5-"4&3
$0/6&8
)5--
0'
$)"24
1*$,3
$&/4&2
0'
(2"6*49
!
$)"24
7*4)
#*((&2
)5--
"2&"
 
  

5--
&/420*% "-$5-"4&3
$0/6&8
)5--3
1*$,3
-"2(&2

$&/420*%
"-0/(
2&-&6"/4
%*.&/3*0/

0/-9
2"1
2&" 2"73
42"1&:0*%3
#&47&&/
&"$)

3)"1&<3
&84&2/"-
#05/%"2*&3
$"/
#&

$)"24<3
401
"/%
#0440.
#"23
1*$,3
.02&
6*35"--9
3"-*&/4
4)"/
*/4&2/"-

#*((&2
"2&"
#05/%"2*&3
!
7)*$)
$05-%
120%5$&

06&27&*()4*/(
0'
4)&
'*234
"/%
-"34
#"23

2"1
&/420*% 2"73
42"1&:0*%3
#&47&&/
&"$)
7)&/
+5%(*/(
4)&
#05/%"29
$0/4052

$)"24<3
401
"/%
#0440.
#"23
1*$,3

42"1&:0*%
7*4)
-"2(&2
$&/420*%

9..&429
*"3 "-$5-"4&3
3,&7
*&
39..&429
0'
&01-&
"2&
3&/3*4*6&
40
39..&429
!

&"$)
3&4
$)003&3
7)*$)
$)"24
*3
-&33
"/%
"2&
#*"3&%
40
3&-&$4
39..&42*$

3,&7&%
*&
.02&
39..&42*$ 0#+&$43
&6&/
7)&/
4)&
4"3,
*3
/04
"

39..&429
+5%(.&/4
!

"/(& 842"$43
"--
1"*27*3&
%&-4"3
7*4)*/
"
&/(4)
%*''&2&/4*"4*0/
)"3
)*()
"$5*49

3&4
$)003&3
3&4
7*4)
4)&
-0/(&34
!
1"*27*3&
-&/(4)
%*''&2&/$&

*((&34
06&2
&47&&/
$)"243
'*/%3
-"2(&34
%&-4"

"*2
#3 #&47&&/
*4&.
1"*23
" # 
" # ;

1*$,3
$)"24
7*4)
-"2(&34
103*4*6&
%&-4"
 
  

*((&34
06&2
".&
"3
*((&34
06&2
"*27*3&

"*2
&- #54
3$"-&%
2&-"4*6&
40
4)&
3."--&2
*4&.

7*4)*/
4)&
1"*2
*((&34
0.1"2&
401
*4&.3
1*$,3
$)"24
7*4)
)&/
'"$&%
7*4)
.5-4*1-&
0#+&$43
1&01-&

*234
4&. -"2(&2
401
*4&. "2&
#*"3&%
40
"44&/%
40
4)&
401
0#+&$4

!
*((&34
*%%-&
0.1"2&
.*%%-&
*4&.3
1*$,3
$)"24
&01-&
.*()4
3&-&$4
"
(2051<3
$&/4&2
*4&.

4&. 7*4)
-"2(&2
.*%%-&
*4&. !
"/%
4)&/
3&-&$4
4)&
-"2(&2
0'
4)&
470

!
-01&
*/ */%3
&"$)
$)"24<3
.*/."8
&01-&
.*()4
3&-&$4
.*/."8
054-*&23

40
"8 $0.154&3
3-01&
#&47&&/
4)&.
*$,3
!
4)&/
$"-$5-"4&
0''3&4
!
4)&
$)"24
7*4)
4)&
-&"346&24*$"-
3-01&

Fig. 5: A set of candidate perceptual proxies that might be used in visual comparison of means and ranges (and possibly other tasks). The
proxies are arranged by their correspondence with hypothesized distinctions between global and local visual scopes.

7.1.1 Global Features
Global-level features describe properties aggregated over a visual
set of items, rather than comparing two focal items. Viewers can
rapidly compute global statistics such as the mean of a collection of
items [1, 11, 23, 25], though from present work it is unclear if this
ability is mediated by a proxy. One high-precision proxy is that the
lengths of bars in a set are veridically averaged together and the chart
with the largest ensemble length is chosen as the answer for the task.
The mean length feature tests this veridical averaging. Viewers might
also perceptually organize the bars into a coherent object, such that
what they perceive is the convex hull of the bounded object that in-
cludes the heights of the bars and the white space between bars, and
then compare the centroids or areas of these two hulls. These object
boundary proxies might be subject to perceptual biases, such as over-
weighting outer edges in contour judgments [6]. Empirical research
on human attention suggests that the allocation of attention through-
1018 IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS, VOL. 26,
out visual displays is preceded by the organization of the scene into
objects and groups [5], and that the center-of-area of those objects
can be rapidly computed [17]. The hull area and hull centroid proxies
test whether this visual feature is consistent with participant responses
and consistent with differences in the data. Note that for superposed
charts, the two hulls are overlapping, such that this particular visual
feature may be harder for people to see because it involves filtering
using color rather than space (as with the stacked, mirrored, and verti-
cal arrangements). Finally, people are highly sensitive to symmetry in
displays [24] and are biased to select symmetric over asymmetric in-
formation [15]. One possible heuristic is that people use symmetry as
a proxy for range, such that any chart that is less symmetric is selected
as the one having the bigger range.
We suspected that visual features describing global, rather than
focal, characteristics of the visualization would be better predictors
of human decisions in the M AX M EAN task, because this task in-
volves set-level comparisons. Conversely, because precision for a
M AX R ANGE task requires the isolation of individual item lengths for
comparison between sets, we predicted visual features describing focal
characteristics would be better predictors of human decisions.
7.1.2 Focal Features
Focal features describe pairwise differences between two items. Peo-
ple can discriminate small differences in line segment lengths [16].
Chart viewers might be sensitive to the deltas, either between charts
(Biggest Mover Pair) or within a chart (Neighbor Delta). In addition,
focal attention can be biased to attend to the topmost item in a col-
lection [26], so one possible proxy is that people compare only the
lengths of the topmost items of the two sets (Biggest First Item).
7.1.3 Implementations
We implemented these global and focal perceptual proxies for all
charts (Figure 5 and pseudocode in Appendix B in the supplemental
materials). Some visual features may be salient [13] to human ob-
servers, but not useful for an analytic task (uncorrelated with the an-
swer). For example, the delta between adjacent bars (i.e., the amount
of overhang) might be a salient and useful indicator for an analytic
task involving comparing items, but if the viewer’s goal is to compare
means, relying on this feature should impair task performance.
To evaluate these proxies, we simulated what would happen if each
proxy was tested on every data series combination that each observer
actually saw in the two experiments. Each proxy was used to make a
decision about a visual comparison (e.g., Hull Area generated a convex
hull around each of the two charts, calculated their areas, and evaluated
the pixel difference in their areas), and provided an “answer” to the
task (i.e., larger area is used as a proxy for mean or for range).
Note that this procedure necessarily shows the proxies different
stimuli depending on arrangement: because the stimuli have been
titrated to respond to viewer accuracy, the charts “shown” for stacked
stimuli will have different properties than the charts “shown” for su-
perposed stimuli. Because the data in the charts “shown” to the proxies
is arrangement-specific, proxies were implemented to be arrangement-
invariant. The proxies were calculated using raw data values, the
length of each mark, and the relative location of each mark (e.g., the
first datum in a chart was at the “top” location), not as visual features
extracted from an image-based representation. Future work should
also test proxy performance using image-based implementations.
We computed two outputs for each of these proxies: which chart
would the proxy have chosen, and was this choice correct? Although
we excluded some participants from the titer analysis for low accu-
racy, we included their data in the simulation to allow for the future
possibility of testing whether their poorer task performance is consis-
tent with using different perceptual proxies than other viewers with
higher-precision visual comparison.
Files that contain trial-by-trial data for properties of the stimuli, hu-
man responses, the pixel information used by each perceptual proxy
to inform a heuristic about a chart decision, and each proxy’s de-
cision, for all combinations of arrangement and task, are posted at
https://osf.io/uenzd/.
NO. 1, JANUARY 2020
7.2 Perceptual Proxy Results
The goal of this proxy approach is to evaluate which visual features
are consistent with human performance, and which are actually useful
for the task. As such we evaluate the “decisions” of each proxy against
two baselines. On what proportion of trials did the proxy agree with
the participant’s response? And on what proportion of trials did the
proxy agree with the true answer of the stimulus? We treat all of the
following results as initial speculations, and make no claims of their
statistical reliability. These values are depicted in Fig. 6. A visual
feature can be considered useful if a decision using the differences in
that visual feature is consistent with the task-dependent differences in
the data. The dots in Fig. 6 to the right of 50% show features that give
above-chance performance at the task. We highlight a few patterns.
First, the most useful visual feature depends on comparison task.
Broadly speaking, global visual features such as centroids are better
candidates for the M AX M EAN comparisons, and focal visual features
such as Neighbor Delta are better candidates for M AX R ANGE compar-
isons. For the M AX M EAN task, visual features of the Mean lengths
(global), Bar Centroids (global), and Biggest Mover Pair (focal) were
the most predictive of the difference in the means. It was unexpected
that the Biggest Mover Pair, which computes pairwise differences be-
tween chart items, predicted the difference of means at above-chance
levels. It suggests that in the data, the largest between-item change
(neighbor delta in superposed charts, motion in animated charts) was
predictive of the chart means, moreso than other global features. For
M AX R ANGE, the Range proxy (which computed all pairwise dis-
tances between items) was most useful, closely followed by pairwise
differences only between neighboring items (Neighbor Delta).
Second, people tend to make decisions consistent with using the
most useful visual features: the bars that show agreement between
proxy responses and human responses tend to follow the dots that
shows the most task-relevant useful features in Figure 6.
Third, we note the absence of a symmetry bias. The Symmetry
proxy, which uses stimulus symmetry as a proxy on which to make
M AX M EAN and M AX R ANGE decisions, was predictive neither of ac-
tual differences in means or range, nor of human responses.
Fourth, there is weak evidence of a bias for people to perform the
M AX R ANGE task with the global proxies of Hull Centroid and/or Area
Trapezoid Centroid, to a higher degree than is actually useful in the
task: note where in Figure 6 the human behavior bars are to the right
of the proxy dots.
We speculate that these findings are broadly consistent with the idea
that global visual features are useful for set-level visual comparisons,
and local visual features are useful for item-level visual comparisons.
M AX M EAN and M AX R ANGE tasks benefit from the same chart ar-
rangements, but use different emergent visual features in these chart
arrangements for visual comparison. Visual comparison is afforded
by more than precision of marks and their arrangements. The “visual”
component of visual comparison may rely on a flexible suite of visual
proxies that viewers can rely on to accomplish a given task, depending
on what visual features are present. The slight bias to erroneously use
global features for the M AX R ANGE task raises the speculative possi-
bility that, in some tasks and arrangements, viewers use global shape-
based proxies even when these proxies are not useful.
7.3 Caveats, Limitations, and Future Directions
We enumerate several important caveats below and how they suggest
possible future avenues of research.
1. Whither the cube? We began this work with the intent of filling
out more cells of the mark × arrangement × task “cube” in Fig-
ure 2. While we quickly found that this model does not seem to
scale, it is possible that more data (varying marks, arrangements,
and tasks) might show that it can scale with some modifications.
For example, dividing the tasks axis into “local” and “global”
might allow a useful level of predictive validity, even if the un-
derlying perceptual explanations are less satisfying.
2. More data on the cells of the cube: Existing work has only
now empirically evaluated four visual comparison tasks, and
JARDINE ET AL.: THE PERCEPTUAL PROXIES OF VISUAL COMPARISON 1019

             

   
  
   
!! 
    !
!!  

  

 
  ! " 
  # " 
!  
$ %& ! 
' 
 
' ' 
$ %& ! 
!  
!!  
!! 
 

  
  

 
 
  # " 
   
  ! " 
    !

Fig. 6: Results of the two analyses of visual feature performance, split by task: M AX M EAN and M AX R ANGE. The x-axis is the percentage of
trials for which the visual proxy was predictive, for human behavior (vertical bars), and for true answer for the comparison (colored dots). The
small dots show individual subjects, and light gray around the black lines shows 95% confidence interval. True answer dots are color-coded to
show whether we informally coded them as a global proxy feature (blue) or focal proxy feature (orange). The true answer dots indicates that
some features are more useful than others for a given visual comparison.

there are many more to test. The present experiment focused
only on bar charts, due to their ubiquity in real visualizations, and
combination of position and length encodings. It will clearly be
important to fill out the “cube” by testing other marks, including
lines, orientations, saturations, etc., both to test the robustness
of the cube model and to provide more data for the enterprise
of searching for candidate perceptual proxies for visualization
tasks. The bar charts in the present work and that of Ondov et
al. [20] were horizontally extended, an increasingly common de-
sign [8]. Other variants even of bar charts might reveal the use
of different proxies for comparison.
3. Different dependent measures: The current experiments titrate
the size of the compared difference, instead of other potential dif-
ficulty manipulations, such as the time allowed to make the judg-
ment, or the number of objects in each set. But some visual com-
parisons need not be precise, and future work should test whether
the same patterns of results hold for these alternative dependant
measures. We would not be surprised by substantial differences
in those results, as the perceptual proxies that help make pre-
cise judgments could differ substantially from those that allow
coarser judgments, or judgments over larger sets of values.
4. Artificial artificial artificial intelligence: We measured the per-
formance of workers on Mechanical Turk (whose slogan is “ar-
tificial artificial intelligence,” because human workers take on
tasks that are often automated) on making visual comparisons,
and then matched their trial-by-trial performance to the predic-
tions of each candidate proxy. Another way to test the match
of the proxies is to create “bots” that perform the same experi-
ments, simulating Mechanical Turk workers that exclusively use
only one proxy. Large numbers of these simulated participants
could run through the real experiment, with difficulty titrated ac-
cording to their performance across trials, in a way that produces
mean titer levels for each “proxy bot,” allowing another type of
comparison of the proxy to human performance.
5. Proxies merely fit performance: This human-proxy agreement
metric can only reveal features that are consistent with human
performance. They cannot confidently reveal what features peo-
ple actually use. As in science more broadly, we can rarely be
sure of an answer, but we can be sure which of many generated
potential answers is most consistent with the data.
6. Proxy overlap: The tested visual features are highly inter-
correlated. As such, we cautiously refrain from strong conclu-
sions about which best predict performance. As above, future
work should use more sophisticated modeling that accounts for
this shared variance. It could also rely on datasets that are in-
tentionally designed to maximally differentiate among the pre-
dictions of the candidate proxies. For example, our data gener-
ation for the M AX M EAN task was specifically designed so that
the largest item was not predictive of the largest set mean. As
such a largest-item proxy could not be useful for the current
M AX M EAN data sets, but could be for other data sets.
7. A proxy for proxies: The proxies implemented here did not use
a computer visual system to “look at” pixels of a chart’s visual
1020 IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS, VOL. 26,
features and parse those pixels into values. We used the actual
data values to generate models of these perceptual proxies. The
value of this approach is that if we can determine the proper-
ties of the data and arrangements that lend themselves to partic-
ular proxies for comparison, then a potential application of this
approach is that an automated visualization system would only
need know the data values and the designer’s desired compari-
son to construct the mark and arrangement to support that com-
parison. In other words, these proxies do not directly take into
account limitations in perceptual visual acuity, or the capacity
limitations of attention and memory.
8. How can one generate new candidate proxies? Future work
should generate more, and more sophisticated, proxies (includ-
ing combinations of proxies, and eventually, predictions for who
will use which, and when). We generated proxies with a com-
bination of intuition and consultation with the perceptual psy-
chology literature, including a strong influence of the literature
on focal vs. global processing modes in vision. Our list is by
no means exhaustive, and identifying new candidates will be a
creative process that, like hypothesis generation across the rest
of science, relies on engaging a diverse group of people with
different types of background knowledge across both the percep-
tion and data visualization communities. A brute force approach
would be to generate the full space of mathematically possible
pairwise and set-wise proxies. Another route could be based on
interviews with viewers engaged in a particular task, to see which
aspects of their proxies might be consciously verbalized.
7.4 Implications for Visualization
The study of the visual features used for visual comparison points to
fruitful paths forward for both perceptual psychology as well as for
inspiring guidelines for effective visual comparison. For practitioners,
because our results indicate that people use different visual features for
different tasks, a visualization designer could use these rules to opti-
mize their visualizations and arrangements depending on the perceived
task. For example, a visualization where the key task is perceiving a
range, such as trend over time in, e.g., a stock market visualization,
should clearly optimize for focal visual comparisons. In visualizations
where understanding the biggest mean is central, value arrangements
should favor global visual features.
Our study here was confined to bar charts, but it is clear from the
richness of our results that this limitation did not restrict the complex-
ity of the performance results. Bar charts are clearly flexible visual
representations in that they support both global and focal visual com-
parison. Nevertheless, another clear next step is to expand this work
to other visual mark types.
In our empirical study of perceptual proxies, the features that
yielded the most similar responses suggests that some proxies are more
likely than others to explain human behavior. This is not to say such
a correspondence would prove that people use these extremely sim-
plified proxies exactly, or alone, but instead point a path forward to
possible mechanisms that can be empirically evaluated.
7.5 How Might Viewers Choose Proxies?
The proxy approach has been a fruitful one for the study of value es-
timation and comparison of correlation in scatterplots [12, 22, 27].
Scatterplots are an ideal “Petri dish” with which to test perceptual
proxies for visualization for a number of reasons. Scatterplots are also
used often to communicate a single statistic (correlation) of a set for
which precision is important. A viewer seeing a scatterplot will likely
develop the analytic goal of perceiving correlation, which should be
more likely to trigger analysis of the proxies available in the scatter-
plot visualization to calculate correlation [21]. Designers are likely to
be implicitly aware of some of these perceptual proxies for global and
focal tasks. It is possible that the ubiquity of bar chart histograms, or
multi-item bar charts in general, is because they allow for both global
and focal comparisons.
NO. 1, JANUARY 2020
One possible avenue of future work is a survey of the literature of
what kinds of visualizations are used for different kinds of visual com-
parisons, to test the validity of the global-versus-focal distinction. A
possible eventual extension of this line of research is the potential au-
tomation of constructing marks and arrangements, based on character-
istics of the data and on the designer’s desired comparison they wish to
enable in the viewer. This line of work would invert the approach used
here. In the current work, we assessed performance of proxies and of
observers with a known task. The experimental task may have guided
the proxy that observers used. But what if a viewer does not have a
specific task before seeing a visualization? One possibility is that the
proxy that yields the most salient comparison compels the viewer to
perform a task guided with that proxy. For example, in a visualiza-
tion with two charts with one large item outlier, the single item out-
lier might capture attention provoke a focal comparison. Future work
could manipulate the salience of different proxies in a stimulus and
evaluate which task observers perform based on the salience of those
proxies, and verify this approach with interviews.
8 C ONCLUSION
In this work, we found evidence that visual comparisons of biggest
mean and biggest range are supported by some chart arrangements
more than others, and that this pattern is substantially different from
the pattern for other tasks. We proposed a series of potential prox-
ies across different tasks, marks, and spatial arrangements. Simple
models of these proxies can be empirically evaluated for their explana-
tory power by comparing their performance to human performance for
these marks, arrangements, and tasks. We used this process to high-
light candidates of perceptual proxies that might scale more broadly to
explain performance in visual comparison.
ACKNOWLEDGMENTS
Nicole Jardine was supported by the U.S. National Science Founda-
tion grant number DRL-1661264 while affiliated with Northwestern
University. Brian Ondov was supported by the Intramural Research
Program of the National Human Genome Research Institute, a part of
the U.S. National Institutes of Health. Any opinions, findings, and
conclusions or recommendations expressed in this material are those
of the authors and do not necessarily reflect the views of the respective
funding agencies.
R EFERENCES
[1] G. A. Alvarez. Representing multiple objects as an ensemble enhances
visual cognition. Trends in Cognitive Sciences, 15(3):122–131, 2011. doi:
10.1016/j.tics.2011.01.003
[2] R. Amar, J. Eagan, and J. Stasko. Low-level components of analytic activ-
ity in information visualization. In Proceedings of the IEEE Symposium
on Information Visualization, pp. 111–117. IEEE, Piscataway, NJ, USA,
2005. doi: 10.1109/INFOVIS.2005.24
[3] J. Bertin. Semiology of Graphics. University of Wisconsin Press, Madi-
son, Wisconsin, 1983.
[4] M. Brehmer, J. Ng, K. Tate, and T. Munzner. Matches, mismatches, and
methods: Multiple-view workflows for energy portfolio analysis. IEEE
Transactions on Visualization and Computer Graphics, 22(1):449–458,
2016. doi: 10.1109/TVCG.2015.2466971
[5] Z. Chen. Object-based attention: A tutorial review. Attention, Perception,
& Psychophysics, 74(5):784–802, 2012. doi: 10.3758/s13414-012-0322
-z
[6] H. Choo, B. R. Levinthal, and S. L. Franconeri. Average orientation is
more accessible through object boundaries than surface features. Jour-
nal of Experimental Psychology: Human Perception and Performance,
38(3):585, 2012.
[7] P. Dragicevic, A. Bezerianos, W. Javed, N. Elmqvist, and J.-D. Fekete.
Temporal distortion for animated transitions. In Proceedings of the ACM
Conference on Human Factors in Computing Systems, pp. 2009–2018.
ACM, New York, NY, USA, 2011. doi: 10.1145/1978942.1979233
[8] S. Few. Now You See It: Simple Visualization Techniques for Quantitative
Analysis. Analytics Press, Oakland, CA, USA, 2009.
JARDINE ET AL.: THE PERCEPTUAL PROXIES OF VISUAL COMPARISON
[9] M. Gleicher. Considerations for visualizing comparison. IEEE Trans-
actions on Visualization and Computer Graphics, 24(1):413–423, 2017.
doi: 10.1109/TVCG.2017.2744199
[10] M. Gleicher, D. Albers, R. Walker, I. Jusufi, C. D. Hansen, and
J. C. Roberts. Visual comparison for information visualization. In-
formation Visualization, 10(4):289–309, Oct. 2011. doi: 10.1177/
1473871611416549
[11] J. Haberman and D. Whitney. Ensemble perception: Summarizing the
scene and broadening the limits of visual processing. From Perception to
Consciousness: Searching with Anne Treisman, pp. 339–349, 2012. doi:
10.1093/acprof:osobl/9780199734337.003.0030
[12] L. Harrison, F. Yang, S. Franconeri, and R. Chang. Ranking visualiza-
tions of correlation using Weber’s law. IEEE Transactions on Visual-
ization and Computer Graphics, 20(12):1943–1952, 2014. doi: 10.1109/
TVCG.2014.2346979
[13] L. Itti, C. Koch, and E. Niebur. A model of saliency-based visual atten-
tion for rapid scene analysis. IEEE Transactions on Pattern Analysis &
Machine Intelligence, (11):1254–1259, 1998.
[14] W. Javed, B. McDonnel, and N. Elmqvist. Graphical perception of mul-
tiple time series. IEEE Transactions on Visualization and Computer
Graphics, 16(6):927–934, 2010. doi: 10.1109/TVCG.2010.162
[15] M. King, G. E. Meyer, J. Tangney, and I. Biederman. Shape constancy
and a perceptual bias towards symmetry. Perception & Psychophysics,
19(2):129–136, 1976. doi: 10.3758/BF03204219
[16] D. M. Levi and S. A. Klein. Vernier acuity, crowding and amblyopia. Vi-
sion Research, 25(7):979–991, 1985. doi: 10.1016/0042-6989(85)90208
-1
[17] D. Melcher and E. Kowler. Shapes, surfaces and saccades. Vision Re-
search, 39(17):2929–2946, 1999. doi: 10.1016/S0042-6989(99)00029-2
[18] T. Munzner. Visualization Analysis and Design. A.K. Peters Visualization
Series. CRC Press, Boca Raton, FL, USA, 2014.
[19] D. Navon. Forest before trees: The precedence of global features in visual
perception. Cognitive Psychology, 9(3):353–383, 1977. doi: 10.1016/
0010-0285(77)90012-3
1021
[20] B. D. Ondov, N. Jardine, N. Elmqvist, and S. Franconeri. Face to face:
Evaluating visual comparison. IEEE Transactions on Visualization and
Computer Graphics, 25(1):861–871, 2019. doi: 10.1109/TVCG.2018.
2864884
[21] S. Pinker. A theory of graph comprehension. In R. Freedle, ed., Artificial
Intelligence and the Future of Testing, pp. 73–126. Lawrence Erlbaum
Associates, Inc., Hillsdale, NJ, USA, 1990.
[22] R. A. Rensink and G. Baldridge. The perception of correlation in scat-
terplots. Computer Graphics Forum, 29(3):1203–1210, 2010. doi: 10.
1111/j.1467-8659.2009.01694.x
[23] D. A. Szafir, S. Haroz, M. Gleicher, and S. Franconeri. Four types of
ensemble coding in data visualizations. Journal of Vision, 16(5):11–11,
2016. doi: 10.1167/16.5.11
[24] J. Wagemans. Characteristics and models of human symmetry detection.
Trends in Cognitive Sciences, 1(9):346–352, 1997. doi: 10.1016/S1364
-6613(97)01105-4
[25] D. Whitney and A. Yamanashi Leib. Ensemble perception. Annual
Review of Psychology, 69:105–129, 2018. doi: 10.1146/annurev-psych
-010416-044232
[26] Y. Xu and S. L. Franconeri. Capacity for visual features in mental ro-
tation. Psychological Science, 26(8):1241–1251, 2015. doi: 10.1177/
0956797615585002
[27] F. Yang, L. Harrison, R. A. Rensink, S. Franconeri, and R. Chang. Cor-
relation judgment and visualization features: A comparative study. IEEE
Transactions on Visualization and Computer Graphics, 25(3):1474–
1488, Mar. 2018. doi: 10.1109/TVCG.2018.2810918
[28] J. S. Yi, Y. ah Kang, J. T. Stasko, and J. A. Jacko. Toward a deeper un-
derstanding of the role of interaction in information visualization. IEEE
Transactions on Visualization and Computer Graphics, 13(6):1224–
1231, 2007. doi: 10.1109/TVCG.2007.70515
[29] J. Zacks and B. Tversky. Bars and lines: A study of graphic commu-
nication. Memory & Cognition, 27(6):1073–1079, 1999. doi: 10.3758/
BF03201236